在线时间:8:00-16:00
迪恩网络APP
随时随地掌握行业动态
扫描二维码
关注迪恩网络微信公众号
开源软件名称:dgl开源软件地址:https://gitee.com/gapyanpeng/dgl开源软件介绍:Deep Graph Library (DGL)Documentation (Latest | Stable) | DGL at a glance | Model Tutorials | Discussion Forum DGL is an easy-to-use, high performance and scalable Python package for deep learning on graphs. DGL is framework agnostic, meaning if a deep graph model is a component of an end-to-end application, the rest of the logics can be implemented in any major frameworks, such as PyTorch, Apache MXNet or TensorFlow. DGL News03/31/2020: The new v0.4.3 release includes official TensorFlow support, with 15 popular GNN modules. DGL-KE and DGL-LifeSci, two packages for knowledge graph embedding and chemi- and bio-informatics respectively, have graduated as standalone packages and can be installed by pip and conda. The new release provides full support of graph sampling on heterogeneous graphs, with multi-GPU acceleration. See our new feature walkthrough and release note. 03/02/2020: Check out this cool paper: Benchmarking Graph Neural Networks! It includes a DGL-based benchmark framework for novel medium-scale graph datasets, covering mathematical modeling, computer vision, chemistry and combinatorial problems. See repo here. Using DGLA data scientist may want to apply a pre-trained model to your data right away. For this you can use DGL's Application packages, formally Model Zoo. Application packages are developed for domain applications, as is the case for DGL-LifeScience. We will soon add model zoo for knowledge graph embedding learning and recommender systems. Here's how you will use a pretrained model: from dgllife.data import Tox21from dgllife.model import load_pretrainedfrom dgllife.utils import smiles_to_bigraph, CanonicalAtomFeaturizerdataset = Tox21(smiles_to_bigraph, CanonicalAtomFeaturizer())model = load_pretrained('GCN_Tox21') # Pretrained model loadedmodel.eval()smiles, g, label, mask = dataset[0]feats = g.ndata.pop('h')label_pred = model(g, feats)print(smiles) # CCOc1ccc2nc(S(N)(=O)=O)sc2c1print(label_pred[:, mask != 0]) # Mask non-existing labels# tensor([[ 1.4190, -0.1820, 1.2974, 1.4416, 0.6914, # 2.0957, 0.5919, 0.7715, 1.7273, 0.2070]]) Further reading: DGL is released as a managed service on AWS SageMaker, see the medium posts for an easy trip to DGL on SageMaker(part1 and part2). Researchers can start from the growing list of models implemented in DGL. Developing new models does not mean that you have to start from scratch. Instead, you can reuse many pre-built modules. Here is how to get a standard two-layer graph convolutional model with a pre-built GraphConv module: from dgl.nn.pytorch import GraphConvimport torch.nn.functional as F# build a two-layer GCN with ReLU as the activation in betweenclass GCN(nn.Module): def __init__(self, in_feats, h_feats, num_classes): super(GCN, self).__init__() self.gcn_layer1 = GraphConv(in_feats, h_feats) self.gcn_layer2 = GraphConv(h_feats, num_classes) def forward(self, graph, inputs): h = self.gcn_layer1(graph, inputs) h = F.relu(h) h = self.gcn_layer2(graph, h) return h Next level down, you may want to innovate your own module. DGL offers a succinct message-passing interface (see tutorial here). Here is how Graph Attention Network (GAT) is implemented (complete codes). Of course, you can also find GAT as a module GATConv: import torch.nn as nnimport torch.nn.functional as F# Define a GAT layerclass GATLayer(nn.Module): def __init__(self, in_feats, out_feats): super(GATLayer, self).__init__() self.linear_func = nn.Linear(in_feats, out_feats, bias=False) self.attention_func = nn.Linear(2 * out_feats, 1, bias=False) def edge_attention(self, edges): concat_z = torch.cat([edges.src['z'], edges.dst['z']], dim=1) src_e = self.attention_func(concat_z) src_e = F.leaky_relu(src_e) return {'e': src_e} def message_func(self, edges): return {'z': edges.src['z'], 'e':edges.data['e']} def reduce_func(self, nodes): a = F.softmax(nodes.mailbox['e'], dim=1) h = torch.sum(a * nodes.mailbox['z'], dim=1) return {'h': h} def forward(self, graph, h): z = self.linear_func(h) graph.ndata['z'] = z graph.apply_edges(self.edge_attention) graph.update_all(self.message_func, self.reduce_func) return graph.ndata.pop('h') Performance and ScalabilityMicrobenchmark on speed and memory usage: While leaving tensor and autograd functions to backend frameworks (e.g. PyTorch, MXNet, and TensorFlow), DGL aggressively optimizes storage and computation with its own kernels. Here's a comparison to another popular package -- PyTorch Geometric (PyG). The short story is that raw speed is similar, but DGL has much better memory management.
Table: Training time(in seconds) for 200 epochs and memory consumption(GB) Here is another comparison of DGL on TensorFlow backend with other TF-based GNN tools (training time in seconds for one epoch):
High memory utilization allows DGL to push the limit of single-GPU performance, as seen in below images. Scalability: DGL has fully leveraged multiple GPUs in both one machine and clusters for increasing training speed, and has better performance than alternatives, as seen in below images.
Further reading: Detailed comparison of DGL and other Graph alternatives can be found here. DGL Models and ApplicationsDGL for researchOverall there are 30+ models implemented by using DGL: DGL for domain applications
DGL for NLP/CV problemsWe are currently in Beta stage. More features and improvements are coming. InstallationDGL should work on
DGL requires Python 3.5 or later. Right now, DGL works on PyTorch 1.2.0+, MXNet 1.5.1+, and TensorFlow 2.1.0+. Using anacondaconda install -c dglteam dgl # cpu versionconda install -c dglteam dgl-cuda9.0 # CUDA 9.0conda install -c dglteam dgl-cuda9.2 # CUDA 9.2conda install -c dglteam dgl-cuda10.0 # CUDA 10.0conda install -c dglteam dgl-cuda10.1 # CUDA 10.1 Using pip
Built from source codeRefer to the guide here. DGL Major Releases
New to Deep Learning and Graph Deep Learning?Check out the open source book Dive into Deep Learning. For those who are new to graph neural network, please see the basic of DGL. For audience who are looking for more advanced, realistic, and end-to-end examples, please see model tutorials. ContributingPlease let us know if you encounter a bug or have any suggestions by filing an issue. We welcome all contributions from bug fixes to new features and extensions. We expect all contributions discussed in the issue tracker and going through PRs. Please refer to our contribution guide. CiteIf you use DGL in a scientific publication, we would appreciate citations to the following paper: @article{wang2019dgl, title={Deep Graph Library: Towards Efficient and Scalable Deep Learning on Graphs}, url={https://arxiv.org/abs/1909.01315}, author={Wang, Minjie and Yu, Lingfan and Zheng, Da and Gan, Quan and Gai, Yu and Ye, Zihao and Li, Mufei and Zhou, Jinjing and Huang, Qi and Ma, Chao and Huang, Ziyue and Guo, Qipeng and Zhang, Hao and Lin, Haibin and Zhao, Junbo and Li, Jinyang and Smola, Alexander J and Zhang, Zheng}, journal={ICLR Workshop on Representation Learning on Graphs and Manifolds}, year={2019}} The TeamDGL is developed and maintained by NYU, NYU Shanghai, AWS Shanghai AI Lab, and AWS MXNet Science Team. LicenseDGL uses Apache License 2.0. |
请发表评论